In the previous article, we discussed B2A infrastructure — making your product data machine-consumable for AI agents. But being readable is only the first step. The more critical question is: after AI reads you, what actually happens?
This question points to an industry-level structural blank.
How an AI-Driven Order Actually Happens
Let us trace a complete AI-driven e-commerce transaction:
- User asks ChatGPT: "Recommend waterproof running shoes under $200"
- ChatGPT queries Shopify and Google Merchant Center product catalogs via UCP
- ChatGPT synthesizes information and recommends three brands
- User clicks one recommendation link and arrives at Brand A's website
- User browses product pages, reads reviews, compares sizing charts
- User adds to cart
- User completes checkout, pays $189
- Shopify records the order
Now the question: who can see this entire chain?
Nobody. ChatGPT knows what it recommended but not whether the user purchased. Shopify knows what the user bought but not what AI recommended. GA4 knows where the user arrived from but its AI source identification covers only approximately 1% of AI activity. UCP knows what product catalogs the agent queried but not what the user did on the website.
This is the evidence gap: the entire measurement chain from AI recommendation to business outcome belongs to no single participant.
Why the Gap Exists
This gap is not a design oversight — it is determined by Agentic Commerce's architecture.
Protocols address machine-to-machine interoperability, not measurement. UCP solves "how can an AI agent safely query a merchant catalog." ACP solves "how can an AI agent safely execute checkout." AP2 solves "how can an AI agent prove it is authorized to spend money." These are machine-machine interaction standards that do not concern themselves with human behavior on websites.
AI platforms do not track outbound traffic. ChatGPT does not know (and does not care) whether users who click recommendation links eventually purchase. Perplexity does not know whether citation-link clickers convert. Unlike traditional search engines — where Google can at least track click-to-conversion via Ads and GA4 — AI recommendation platforms currently have no similar feedback mechanism.
Merchant analytics tools do not understand AI upstream. GA4 can identify some AI-sourced referrers (it added the "AI Assistants" channel group in March 2026), but it cannot see AI recommendation context — what the user asked, which competitors AI recommended, why AI recommended you. It certainly cannot see agent server-side requests — BrightEdge data shows these account for 15% of total traffic, entirely within GA4's blind spot.
Commercial Consequences of the Gap
This evidence gap is producing real consequences at three levels:
First, brands cannot attribute AI revenue. Shopify says AI orders grew 13x. But can you, as a Shopify brand, determine how much of your revenue comes from AI? Most brands answer "uncertain" — because GA4's AI source identification is incomplete, Shopify orders lack AI source labels, and no session-to-order join connects AI arrivals with Shopify orders.
Second, brands cannot optimize the AI channel. Traditional paid search has clear optimization levers at every funnel step, each with data and A/B testing capability. For the AI channel, you do not know what reasoning AI gave when recommending you, which competitors it presented alongside you, how AI-referred users behave differently from other channels, or even whether traffic labeled "Direct" actually came from AI.
Third, AI gets undervalued in resource allocation. When a CMO allocates budget among Google Ads, Meta Ads, Email, and "AI/GEO," the first three have clear ROAS data. AI/GEO has "our AI traffic seems to be growing, but we are not sure exactly how much revenue it drives." In this information asymmetry, AI almost inevitably loses to traditional channels with more complete data — not because AI is actually less efficient, but because it lacks measurement infrastructure.
CitationGraph's Structural Opportunity
If we map each participant's coverage onto the transaction chain:
[UCP Coverage] [ACP Coverage]
AI queries catalog ──→ AI recommends ──→ User clicks ──→ Browse ──→ Cart ──→ Checkout ──→ Order
|←─────────── Evidence Gap ──────────→|
|←──── CitationGraph Coverage ────→|CitationGraph's capabilities precisely correspond to this evidence gap. Answer Evidence (SOV/Prompt Sampling) covers the AI recommendation stage. Request Visibility (Edge Lite/Log Bridge) covers agent data fetching. First-party JS and AI referrer identification cover user arrival. Shopify Web Pixel integration covers on-site commercial behavior. Session-to-order join covers revenue attribution.
This is not accidental product positioning. It is the result of starting from "AI readability diagnostics" and extending simultaneously upstream (Answer/Request) and downstream (Commerce/Attribution) along the AI commercial chain. GEO tools cover only upstream. Attribution tools cover only downstream. CitationGraph is the only platform building full-chain coverage.
Protocols Will Not Automatically Fill the Gap
One might think: once UCP and ACP mature, will this gap naturally disappear? No, for three reasons.
First, protocols solve interoperability, not measurement. UCP enables catalog queries but does not measure whether queries produce recommendations. ACP enables checkout but does not measure what happened before checkout. Protocols are pipes, not monitoring systems.
Second, the gap involves human behavior. From a user clicking an AI recommendation link to completing a purchase, what happens in between is human browsing, comparing, hesitating, abandoning, or buying on a website. Only the merchant's own tracking systems can observe these behaviors — no external protocol can cover them.
Third, cross-platform fragmentation will only intensify. Users may research on Perplexity, compare on Gemini, confirm on ChatGPT, then type the URL directly. Each AI platform sees only its own segment. Merchants need a cross-platform evidence aggregation layer.
AIAA Is the Framework for Filling the Gap
The AIAA five-layer framework maps directly onto the evidence gap: Answer layer covers AI recommendation, Request layer covers agent data fetching, Visit layer covers user arrival, Commerce layer covers on-site commercial behavior, Attribution layer covers revenue traceability. AIAA does not replace protocols — it fills the measurement vacuum between them.
Practical Recommendations for Brands
Do not wait for protocol maturity to start measuring AI impact — full protocol interoperability may take 18-36 months, but AI's impact on your business is happening today.
Start with the Visit layer: deploy first-party JS to identify AI-sourced visits. Then deploy Edge Lite to see agent server-side requests. Finally, build a session-to-order join for the Attribution layer.
What Comes Next
The evidence gap's existence means traditional attribution fails in the AI era. In the next article, we analyze the structural crisis of AI attribution — why GA4's last-click model systematically undervalues AI's contribution, why AppsFlyer cannot see AI upstream, and how a layered attribution model solves this problem.
FAQ
Q1: What is the "evidence gap" in the AI commercial chain?
A: The entire measurement chain from AI recommendation (Answer layer) to business outcomes (orders/revenue). UCP covers the discovery end (product queries), ACP covers the execution end (checkout), but the middle — AI recommendation context, user clicks, on-site browsing, add-to-cart behavior — is not observed by any protocol. Brands need an independent measurement layer to fill this gap.
Q2: Why won't the gap disappear when protocols mature?
A: Three reasons: (1) Protocols solve machine interoperability, not measurement. (2) The gap involves human behavior observable only by merchant tracking systems. (3) Cross-platform fragmentation will intensify as users switch between multiple AI platforms.
Q3: How does CitationGraph fill this gap?
A: Through the AIAA five-layer framework covering every gap segment: Answer (AI recommendation), Request (agent data fetching), Visit (user arrival), Commerce (on-site behavior), Attribution (traceable revenue). CitationGraph is the only platform building full-chain coverage from AI readability to business outcomes.
Q4: Which part of the gap should brands fill first?
A: Start with the Visit layer — deploy first-party JS to identify AI-sourced visits. Then Edge Lite (agent server-side requests). Then session-to-order join (Attribution layer). You do not need to cover all five layers simultaneously.
Q5: What does this gap mean for advertising teams?
A: AI gets undervalued in budget competition. Google Ads and Meta Ads have complete ROAS data; AI/GEO does not. Without filling the evidence gap, CMOs will reallocate resources from AI to traditional channels with more complete data — even if AI's actual efficiency is higher.